Physical activity (PA) can improve physical and psychosocial well-being among adults diagnosed with cancer. Benefits reported throughout the cancer trajectory (ie, from diagnosis onward) include enhanced physical functioning and quality of life, as well as reduced negative effects of cancer and treatment-related side effects [1]. Consequently, cancer-specific PA guidelines have been published, recommending at least 90 minutes of weekly moderate-intensity aerobic PA (note: before 2019, 150 minutes were recommended) and strength training for ≥2 days each week [2,3]. These guidelines have also been endorsed by leading cancer support organizations [4]. Despite this evidence, most adults diagnosed with cancer do not achieve the recommended PA levels [5].

Thus, developing and testing interventions to increase PA levels is a priority. As described in recent systematic reviews and meta-analyses, most interventions designed to enhance PA levels among individuals with cancer have been delivered face-to-face in fitness facilities, and findings suggest that such interventions can enhance physical and psychosocial well-being [6]. However, among adults diagnosed with cancer, barriers such as lack of time, limited access to facilities, and travel distances can hinder participation in face-to-face PA interventions [7]. Barriers to PA have been exacerbated during the COVID-19 pandemic, with most face-to-face PA opportunities being limited or canceled and adults with cancer reporting decreased PA and increased sedentary time [8].

eHealth technologies, including telephones, websites, email, and mobile health (mHealth) technologies (eg, SMS text messaging, smartphones, wearable technology, and apps) may be useful to address some of these barriers to PA and reach a wider audience of adults living with cancer [9-11]. The prevalence of and preference for using eHealth is increasing rapidly among adults with cancer, with the National Cancer Institute prioritizing research into the effective use of eHealth in the context of PA promotion for adults with cancer [12-14]. Reviews summarizing the effects of eHealth to promote PA in adults with cancer suggest that technology-supported PA interventions may enhance PA levels and health-related quality of life and decrease fatigue [15-19]. Notwithstanding the evidence to date, important gaps in knowledge remain. First, only studies published before July 2018 have been reviewed. As the field of eHealth PA interventions is rapidly growing and evolving, an update is needed. Second, reviews have had limited scope with regard to study design (eg, randomized controlled trials [RCTs] only [18]), population (eg, women with breast cancer only [19]), and technology components (eg, activity trackers or mHealth only [16,17,19]). Expanding eligibility criteria to include various study designs, cancer types, and the full range of eHealth technologies is required to provide a more comprehensive overview of the effects of eHealth PA interventions in oncology. Finally, despite evidence supporting the role of behavior change techniques (BCTs) and theories (eg, theory of planned behavior) in PA interventions, the integration of BCTs and theory with eHealth PA interventions has received limited attention [15,18,20-22]. Roberts et al [15] examined the use of theory and BCTs for 15 eHealth PA interventions published before November 2016, whereas Kiss et al [18] coded BCTs for 16 interventions, many of which were duplicates from Roberts et al [15], published before July 2018.

Thus, the purpose of this review is to summarize evidence on the use of eHealth to support PA behavior change among adults diagnosed with cancer. The specific objectives are to (1) describe the current state of the literature on the effectiveness of eHealth in supporting PA behavior change (pre- to postintervention and follow-ups, where available), (2) explore intervention characteristics that may promote PA behavior change (eg, eHealth components, use of theory, and BCTs), and (3) identify research needs for future work.

For identifying relevant studies, a search strategy covering the major topics of health technology, cancer, and PA was developed in MEDLINE (R) using existing reviews to guide the selection of search terms. It was then refined, finalized, and translated to the other databases used herein with the help of a university librarian (Table S1 in Multimedia Appendix 1). MEDLINE (R) and Epub Ahead of Print, In-Process and Other Non-Indexed Citations and Daily (OVID), Embase (OVID), CENTRAL (OVID), CINAHL (EBSCO), Sport Discus (EBSCO), and Scopus were searched from database inception through to December 18, 2019. This search was updated on January 7, 2021.

To be included, articles had to (1) comprise adult participants aged ≥18 years diagnosed with cancer, (2) evaluate a PA intervention that used technology (mobile app, SMS text messages, wearable activity tracker, website, email, or other eHealth) as an active component in the intervention to support behavior change, (3) measure and report on PA levels (objectively or subjectively), (4) be published in English, and (5) be published in a peer-reviewed journal (conference abstracts and gray literature were not included). Articles were excluded if they (1) involved adults whose only cancer diagnosis occurred during childhood, adults without a history of cancer, or caregivers; (2) used telephone contact as the only technology component in the intervention; (3) used technology for the measurement of outcomes only (eg, accelerometer for PA measurement pre- or postintervention); (4) lacked a PA intervention (eg, observational study of PA behavior); (5) reported ongoing trials without full results being available (ie, protocols); and (6) the full text was unavailable. Interventions could be either partially supervised (ie, some human contact) or unsupervised (ie, entirely automated), and the amount of technology use within interventions was not quantified.

After importing all search results into EndNote X9.2 (Clarivate Analytics), the first author conducted automatic and subsequent manual deduplication. Unique articles were exported to Rayyan (Rayyan Systems) for screening according to the eligibility criteria [24]. Title and abstract screening were conducted concurrently by the first author by removing all articles that did not meet the criteria. Articles with titles and abstracts that lacked enough information to make a decision were carried forward to the full-text screening stage. Full texts of the remaining articles were obtained and screened independently by the first (ME) and second authors (MME), who recorded their decisions as well as reasons for exclusion where applicable. The 2 authors then met to discuss the decisions and resolve disagreements based on additional reviews of the articles. Disagreements that could not be resolved directly were resolved via discussion with the senior author (NCR) to yield the final list of included articles.

Before data extraction, a standardized data extraction table was developed and refined using 3 test articles. The final data extraction table included (1) participant information (age, cancer diagnosis, and eligibility criteria), (2) study design (timing, eligibility and recruitment rates, and recruitment methods), (3) intervention details (groups, objectives, duration, active components, technology integration, BCTs according to the Michie behavior change taxonomy comprising 93 BCTs across 16 categories [25], and use of theory), (4) outcomes (participant numbers, demographics, primary and secondary outcomes, PA-related outcomes, adherence or completion to intervention, and technology use), and (5) additional factors (key findings, challenges, and limitations). It was decided that theory would be recorded only when explicitly described in the included studies. Data were then extracted independently by the first (ME) and second authors (MME), with each author being responsible for half the number of articles. For confirming the reliability of the extraction, 5 random articles were exchanged between authors, extracted a second time, and the data were compared between extractions. Because of minor discrepancies, coding of BCTs was repeated for all articles, and discussions were held between the first and second authors to reach a consensus. The authors did not complete BCT coder training before BCT coding. No other discrepancies were noted. Any missing information was denoted using the phrase not reported in the data extraction table. Attempts were made to fill in missing information via protocol papers and other related publications for each study. The authors of the included articles were not contacted directly for additional information.

The Cochrane risk of bias (RoB) tool (RoB-2) was used for multiarm interventions, which included evaluations for RoB in five domains: (1) randomization, (2) deviation from the intended intervention, (3) missing outcome data, (4) measurement of the outcome, and (5) selection of reported results [26]. The ROBINS-I (RoB in nonrandomized studies of interventions) tool, which evaluates bias across seven domains: (1) confounding, (2) participant selection, (3) classification of intervention, (4) deviation from intended intervention, (5) missing data, (6) outcome measurement, and (7) selection of reported results was used for single-arm designs [27]. An overall RoB was given according to the highest RoB rating in any domain for each study. For example, a study with high RoB in domain 1 and low RoB across all other domains received a high overall RoB rating. The completeness of reporting was evaluated using the CONSORT (Consolidated Standards of Reporting Trials)–eHealth checklist, with items assessed as reported, not reported, or not applicable [28]. The completeness of reporting score was calculated for each article as the percentage of applicable items that were reported. These assessments were performed independently by the first (ME) and second authors (MME). Verification was performed by exchanging 5 random articles between authors for repeat assessment, and no discrepancies were documented.

To summarize the data extracted from each article, descriptive statistics were calculated for participant demographics, adherence, and completion. Intervention details were categorized and summarized, whereas results were converted to standardized metrics where possible to enable comparison across studies. Because of the substantial heterogeneity of the studies with regards to population, intervention, comparison, and outcome, meta-analyses were not performed. Instead, extracted data across studies were summarized using narrative synthesis techniques, and summary tables were presented [29]. Harvest plots were created to provide a visual summary of study effects on PA outcomes, including PA levels directly postintervention and PA maintenance at follow-up, providing an overview of intervention effectiveness on PA levels [30]. Following recommendations, harvest plots were prepared with studies grouped according to the statistical significance of their PA outcomes (PA increase, PA decrease, or no change) [30]. Bar heights were used to distinguish between RCTs (high) and other study designs (low), whereas shading was used to specify how PA was measured (subjective, objective, or both). For addressing objective 2, weight analyses were conducted to explore associations between independent variables (intervention characteristics: use of supervised elements, various types of eHealth, theory, and BCTs) and the dependent variable (PA levels) [31]. Weight was calculated for each independent or dependent variable pair by dividing the number of studies featuring each independent variable and reporting a significant improvement in the dependent variable by the total number of studies featuring the independent variable. Weights range from 0-1, with a higher value indicating a stronger association between the independent variable and significant changes in PA levels. Weights are presented to three decimal places and are equivalent to percentages (ie, 0.123 could also be read as 12.3%). The weight for each independent or dependent variable pairing was then compared with the overall weight for all studies to explore if the presence of certain intervention characteristics was associated with a higher weight (ie, more often linked with significant changes in PA levels). For continuous independent variables (duration and number of BCTs used), studies were grouped according to the mean value (greater than or less than the mean). For BCTs, weights were only calculated for the most common BCTs or BCT categories (ie, used in at least 50% of interventions) to minimize the introduction of further bias when calculating weights using only a small number of independent or dependent variable pairs [31].

After deduplication, 4022 citations were screened at the title or abstract level; of the 4022 citations, 3873 (96.29%) were removed as they did not meet the eligibility criteria. During full-text screening, the agreement between the first 2 authors on the 145 articles was 82.1%, with decisions for articles where no agreement was reached (26/145, 17.9%) being resolved via discussion with the senior author (NCR). Of the 145 articles, 74 (51.0%) articles were excluded during full-text screening, and, overall, 71 (49.0%) articles representing 67 unique studies were included [32-102]. Figure 1 presents an overview of the study selection, with reasons for article exclusion. The remainder of the results are presented according to the number of unique studies (n=67). Tables 1 and 2 provide more information on each of the included studies and their respective PA interventions.

The overall RoB among the 45 RCTs ranged from some risk (4/45, 8%) to high risk (41/45, 91%). This was largely because of RoB in deviation from the intended intervention (7/45, 15% some risk; 38/45, 84% high risk) and measurement of the outcome (31/45, 68% high risk). Most studies had a low RoB for the remaining categories (n=34-44, depending on the category). Because of the risk of confounding, 95% (21/22) of the nonrandomized studies were found to have critical RoB. RoB in the measurement of outcome was moderate (10/67, 15%) or serious (9/67, 13%) for most single-arm studies, whereas it remained low across other categories (see Figure S4 in Multimedia Appendix 1 for RoB among the included studies [32-102]). If not for the lack of blinding, then only 58% of studies would have had a high overall high RoB, mainly because of bias in outcome measurement owing to the reliance on self-reported PA. Mean completeness of reporting was moderate, with 69.4% (71.4% for RCTs and 65.2% for nonrandomized studies) of applicable CONSORT-eHealth items covered in the included publications. Nearly one-third of the applicable items (mean of 30.6%, SD 9.4%) were not reported. For RCTs and nonrandomized studies, mean values of 15.5% (SD 3.4%) and 32.4% (SD 4.7%), respectively, of CONSORT-eHealth items (overall mean 20.8%, SD 8.8%) were not applicable on a case-by-case basis.

The purpose of this review was to provide a comprehensive, updated overview of eHealth intervention research designed to promote PA and to explore intervention characteristics (ie, duration, delivery modalities, use of theory, and BCTs) associated with increased PA levels. Many of the included studies were published after July 2018 and focused on feasibility, which indicates the rapidly growing yet early state of the field. Across the studies, there was substantial heterogeneity in the participants, interventions, and outcomes. All studies had high RoB for some domains, and incomplete reporting was problematic. Nevertheless, findings suggest that eHealth may be an effective strategy to enhance PA levels with selected modalities, BCTs, and behavioral theories that potentially enhance effectiveness.

The growing number of published articles reporting on eHealth PA interventions for adults with cancer (48% of articles published since July 2018) aligns with several funding calls for eHealth research, institutional strategic priorities, and the growing prevalence of, and preference for, eHealth among adults with cancer [12-14]. With the restrictions imposed by the COVID-19 pandemic on face-to-face PA programs, continued acceleration in this field is expected [108]. The COVID-19 pandemic has highlighted the need for eHealth PA interventions in oncology, and such interventions will continue to remain relevant beyond the pandemic, especially for improving the reach of PA interventions to underserved populations with cancer (eg, remote or rural) [8,108]. For example, an ongoing study in Canada that aims to bring exercise oncology programs to remote and rural cancer populations has delivered all classes remotely during the COVID-19 pandemic and will continue to offer videoconference-based programs (NCT04478851) [109,110]. As many of the included studies tested the feasibility of using eHealth for PA promotion in adults with cancer (36%) using single-arm designs or smaller RCTs, the findings on the effectiveness to change PA levels remain largely preliminary. Next steps could include study designs, such as factorial RCTs or alternative trial designs with the capacity to quantify the contribution of intervention effectiveness from various technology components, theories, and BCTs. Finally, larger multisite RCTs or meta-analyses of comparable studies to strengthen the evidence for the effectiveness of these interventions will be required to continue to grow our knowledge [111-113].

Overall, this review highlights that eHealth interventions can increase PA levels, with 52% of the studies reporting significant increases in postintervention PA. Previous reviews have reported that 50%-80% of eHealth PA interventions for adults with cancer reported significant improvements in PA levels [15-19]. Differences in these findings maybe because of the inclusion of studies that were underpowered to detect changes in PA levels (ie, feasibility trials and those aiming to impact a primary outcome other than PA levels), as well as intervention heterogeneity (ie, varied duration, delivery modalities, use of theory, and BCTs). Nevertheless, eHealth PA interventions have the potential to enhance PA levels, although optimization is required. The first step to optimization is to examine eHealth PA intervention components and their impact on effectiveness to change PA behavior.

Findings from this review show that both well-established eHealth components (eg, informational websites) and emerging technologies (eg, mHealth) were associated with increased PA levels both when used alone or in combination with other eHealth. Researchers are encouraged to consider the pros and cons for each type of eHealth when designing eHealth PA interventions. For example, the pros of mHealth include the ability to deliver real-time, context-aware behavior change interventions; passively monitor PA; and relative ubiquity in developed countries (eg, nearly 90% of Canadians own a smartphone) [11,114,115]. Meanwhile, websites that have the highest weight of any eHealth component may be selected for their familiarity and ease of use among older adults [116]. Moving forward, remaining flexible to align eHealth interventions with participant needs and preferences will likely be important [117,118].

A finding from this review that stands in contrast to those of previous reviews in exercise oncology is that a higher percentage of unsupervised interventions (56%; those without face-to-face interaction) were successful at increasing PA levels compared with those that were partially supervised (41%; those with one or more face-to-face components) [7,119]. This may be because of feelings of autonomy promoted by unsupervised interventions, a factor that has been linked to increased intrinsic motivation and PA behavior change [120-122]. In addition, it may be in part because of the more frequent use of behavioral theories (unsupervised: 63%; supervised: 56%) and BCTs (unsupervised mean: 13.8; supervised mean: 11.8) in the included unsupervised interventions, which have been associated with effectiveness in web-based behavioral interventions [123]. Direct comparisons of unsupervised and partially supervised eHealth PA interventions will be required to draw definitive conclusions on their relative effectiveness.

Recommendations have been made to use behavioral theories to guide intervention design to enhance the effectiveness of behavior change interventions [21,22]. Common behavioral theories, such as social cognitive theory, the transtheoretical model, and the theory of planned behavior, have been used in roughly half of eHealth PA interventions for adults with cancer [103-105]. Although the weights for studies using social cognitive theory, the transtheoretical model, the theory of planned behavior, or multiple theories (0.565-0.750) were higher than of those using none at all (0.500), 50% of the interventions that were not theory based also resulted in significant increases in PA levels. Furthermore, it is possible that some articles may have drawn upon theoretically based intervention components without explicitly discussing the use of theory. These mixed results add to the ongoing debate on the role of behavioral theories in real-world interventions [124]. Further examination of the use of theory (eg, theoretical integration and/or use of technology-specific models or theories) is needed to understand its impact, or lack thereof, in eHealth PA interventions.

The most commonly used BCTs in this review of eHealth PA interventions were goal setting and self-monitoring, which is similar to what has been reported in face-to-face PA interventions [20]. However, more BCTs were used across studies in this review, for both mean number per study and overall variety, than in reviews assessing face-to-face interventions [20]. Notably, current findings align with earlier research that has also suggested that certain BCTs may be more effective than others [20,125,126]. Further research is needed to understand the use of BCTs (ie, types and combinations) and their potential impact on intervention effectiveness in eHealth PA research. Indeed, these weight analyses revealed that eHealth interventions with more BCTs were more likely to report significant improvements in PA levels.

Most reviewed studies (93%) had high overall RoB (ie, in one or more domains). This was, in large part, because of the lack of blinding. The inability to blind participants and researchers to PA interventions is a commonly reported limitation, irrespective of eHealth use [18,127]. Consequently, if this domain were removed, then the RoB would remain high in only 58% of the studies, primarily because of the reliance on self-reported PA outcomes [128]. Where possible, researchers may wish to integrate both objective and subjective PA measures into studies to reduce RoB [128]. Objective PA assessment is increasingly accessible, given the activity trackers in mHealth (eg, phones) and decreasing costs. Finally, the finding that all included studies were incompletely reported is problematic. Researchers are urged to follow the reporting guidelines appropriate for their study design, which can be found on the web [129].

There are important considerations to keep in mind when interpreting the findings. The broad inclusion criteria of the review, although selected intentionally to provide a comprehensive overview of this emerging field, hindered the ability to perform quantitative meta-analyses. Despite the systematic review, additional articles may have been missed if published in gray literature or in other languages. Although weight analyses were performed to provide insights for future research, their outputs must be interpreted with caution, as they are not a measure of statistical significance. Any reported associations remain purely exploratory and must be substantiated in future robust study designs. In addition, more than half of the included studies were underpowered to detect changes in PA as a secondary outcome, which is likely to bias weights toward the null. Some study characteristics in the weight analyses were represented in only a few studies, and most studies used complex interventions, making it difficult to identify the effect of individual components on outcomes. Finally, the authors did not complete BCT coder training before extraction, which may have led to some inaccuracies in BCT coding. However, efforts were made to minimize errors by double checking all codes and discussing with the senior author (NCR), an expert in PA behavior change, as needed.

Consolidating the evidence on eHealth PA interventions for adults with cancer led to the identification of several research needs and opportunities that remain to be addressed. First, only 9 studies featured follow-up assessments to track PA behavior change after intervention completion. Examining the long-term maintenance of PA is critical to determine whether these interventions can have a lasting impact on PA levels. Second, it will be important to explore whether completely unsupervised eHealth interventions or eHealth interventions with limited supervision can rival the effectiveness of face-to-face supervised PA programs to increase PA levels in adults with cancer. Such work is needed to advocate for eHealth use in this field and may be crucial to the implementation of scalable PA programs for adults with cancer. Third, examining the effectiveness of videoconferencing platforms, which have surged in popularity during the COVID-19 pandemic, is warranted. Videoconferencing has the potential to leverage the advantages of supervised interventions (eg, live tailored feedback, social interaction, and accountability) while remaining accessible [108]. Fourth, given the rapidly evolving nature of eHealth, testing effectiveness using fully powered alternative trial designs (eg, SMART [sequential multiple assignment randomized trial], microrandomized trials, and factorial RCTs) is warranted so that evaluation can better match the pace of development, heighten external validity, and inform the translation of evidence to practice [112,113]. Such designs also allow researchers to establish definitive links between intervention components and changes in PA levels, allowing for systematic optimization of effectiveness. Finally, evaluations of cost-effectiveness are needed to inform real-world implementations of eHealth PA behavior change programs, as none were reported herein [130].

This review summarizes findings from the rapidly growing field of eHealth PA interventions for adults affected by cancer. Although eHealth use in these interventions varies widely, the results are suggestive of positive outcomes. Furthermore, most studies integrated BCTs and relevant theories. Efforts are required to understand eHealth PA interventions better by exploring the impact on PA maintenance, investigating ways to optimize their effectiveness (by using BCTs, theories, and emerging technologies), and affirming effectiveness by applying well-powered alternative trial designs. Despite the early and evolving nature of this field, positive results suggest there is a case for integrating eHealth with efforts to promote PA, health, and well-being for adults affected by cancer.